Merge pull request #78168 from clayjohn/GLES3-gaussian-rt

Use Gaussian approximation for backbuffer mipmaps in GL Compatibility renderer

Author: YuriSizov

drivers/gles3/effects/copy_effects.cpp

@@ -31,6 +31,7 @@
 #ifdef GLES3_ENABLED
 
 #include "copy_effects.h"
+#include "../storage/texture_storage.h"
 
 using namespace GLES3;
 
@@ -133,6 +134,7 @@ void CopyEffects::copy_screen() {
 	draw_screen_triangle();
 }
 
+// Intended for efficiently mipmapping textures.
 void CopyEffects::bilinear_blur(GLuint p_source_texture, int p_mipmap_count, const Rect2i &p_region) {
 	GLuint framebuffers[2];
 	glGenFramebuffers(2, framebuffers);
@@ -158,6 +160,80 @@ void CopyEffects::bilinear_blur(GLuint p_source_texture, int p_mipmap_count, con
 	glDeleteFramebuffers(2, framebuffers);
 }
 
+// Intended for approximating a gaussian blur. Used for 2D backbuffer mipmaps. Slightly less efficient than bilinear_blur().
+void CopyEffects::gaussian_blur(GLuint p_source_texture, int p_mipmap_count, const Rect2i &p_region, const Size2i &p_size) {
+	GLuint framebuffer;
+	glGenFramebuffers(1, &framebuffer);
+	glBindFramebuffer(GL_FRAMEBUFFER, framebuffer);
+
+	glActiveTexture(GL_TEXTURE0);
+	glBindTexture(GL_TEXTURE_2D, p_source_texture);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+	Size2i base_size = p_size;
+
+	Rect2i source_region = p_region;
+	Rect2i dest_region = p_region;
+
+	Size2 float_size = Size2(p_size);
+	Rect2 normalized_source_region = Rect2(p_region);
+	normalized_source_region.position = normalized_source_region.position / float_size;
+	normalized_source_region.size = normalized_source_region.size / float_size;
+	Rect2 normalized_dest_region = Rect2(p_region);
+	for (int i = 1; i < p_mipmap_count; i++) {
+		dest_region.position.x >>= 1;
+		dest_region.position.y >>= 1;
+		dest_region.size.x = MAX(1, dest_region.size.x >> 1);
+		dest_region.size.y = MAX(1, dest_region.size.y >> 1);
+		base_size.x >>= 1;
+		base_size.y >>= 1;
+
+		glBindTexture(GL_TEXTURE_2D, p_source_texture);
+		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, i - 1);
+		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, i - 1);
+		glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, p_source_texture, i);
+#ifdef DEV_ENABLED
+		GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+		if (status != GL_FRAMEBUFFER_COMPLETE) {
+			WARN_PRINT("Could not bind Gaussian blur framebuffer, status: " + GLES3::TextureStorage::get_singleton()->get_framebuffer_error(status));
+		}
+#endif
+
+		glViewport(0, 0, base_size.x, base_size.y);
+
+		bool success = copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_GAUSSIAN_BLUR);
+		if (!success) {
+			return;
+		}
+
+		float_size = Size2(base_size);
+		normalized_dest_region.position = Size2(dest_region.position) / float_size;
+		normalized_dest_region.size = Size2(dest_region.size) / float_size;
+
+		copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, normalized_dest_region.position.x, normalized_dest_region.position.y, normalized_dest_region.size.x, normalized_dest_region.size.y, copy.shader_version, CopyShaderGLES3::MODE_GAUSSIAN_BLUR);
+		copy.shader.version_set_uniform(CopyShaderGLES3::SOURCE_SECTION, normalized_source_region.position.x, normalized_source_region.position.y, normalized_source_region.size.x, normalized_source_region.size.y, copy.shader_version, CopyShaderGLES3::MODE_GAUSSIAN_BLUR);
+		copy.shader.version_set_uniform(CopyShaderGLES3::PIXEL_SIZE, 1.0 / float_size.x, 1.0 / float_size.y, copy.shader_version, CopyShaderGLES3::MODE_GAUSSIAN_BLUR);
+
+		draw_screen_quad();
+
+		source_region = dest_region;
+		normalized_source_region = normalized_dest_region;
+	}
+	glBindFramebuffer(GL_FRAMEBUFFER, 0);
+	glDeleteFramebuffers(1, &framebuffer);
+
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_LINEAR);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0);
+	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, p_mipmap_count - 1);
+	glBindTexture(GL_TEXTURE_2D, 0);
+
+	glViewport(0, 0, p_size.x, p_size.y);
+}
+
 void CopyEffects::set_color(const Color &p_color, const Rect2i &p_region) {
 	bool success = copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_SIMPLE_COLOR);
 	if (!success) {

drivers/gles3/effects/copy_effects.h

@@ -64,6 +64,7 @@ class CopyEffects {
 	void copy_to_rect(const Rect2 &p_rect);
 	void copy_screen();
 	void bilinear_blur(GLuint p_source_texture, int p_mipmap_count, const Rect2i &p_region);
+	void gaussian_blur(GLuint p_source_texture, int p_mipmap_count, const Rect2i &p_region, const Size2i &p_size);
 	void set_color(const Color &p_color, const Rect2i &p_region);
 	void draw_screen_triangle();
 	void draw_screen_quad();

drivers/gles3/shaders/copy.glsl

@@ -16,17 +16,24 @@ layout(location = 0) in vec2 vertex_attrib;
 out vec2 uv_interp;
 /* clang-format on */
 
-#ifdef USE_COPY_SECTION
+#if defined(USE_COPY_SECTION) || defined(MODE_GAUSSIAN_BLUR)
+// Defined in 0-1 coords.
 uniform highp vec4 copy_section;
 #endif
+#ifdef MODE_GAUSSIAN_BLUR
+uniform highp vec4 source_section;
+#endif
 
 void main() {
 	uv_interp = vertex_attrib * 0.5 + 0.5;
 	gl_Position = vec4(vertex_attrib, 1.0, 1.0);
 
-#ifdef USE_COPY_SECTION
+#if defined(USE_COPY_SECTION) || defined(MODE_GAUSSIAN_BLUR)
 	gl_Position.xy = (copy_section.xy + uv_interp.xy * copy_section.zw) * 2.0 - 1.0;
 #endif
+#ifdef MODE_GAUSSIAN_BLUR
+	uv_interp = source_section.xy + uv_interp * source_section.zw;
+#endif
 }
 
 /* clang-format off */
@@ -39,6 +46,7 @@ uniform vec4 color_in;
 #endif
 
 #ifdef MODE_GAUSSIAN_BLUR
+// Defined in 0-1 coords.
 uniform highp vec2 pixel_size;
 #endif
 
@@ -55,4 +63,31 @@ void main() {
 #ifdef MODE_SIMPLE_COLOR
 	frag_color = color_in;
 #endif
+
+// Efficient box filter from Jimenez: http://www.iryoku.com/next-generation-post-processing-in-call-of-duty-advanced-warfare
+// Approximates a Gaussian in a single pass.
+#ifdef MODE_GAUSSIAN_BLUR
+	vec4 A = textureLod(source, uv_interp + pixel_size * vec2(-1.0, -1.0), 0.0);
+	vec4 B = textureLod(source, uv_interp + pixel_size * vec2(0.0, -1.0), 0.0);
+	vec4 C = textureLod(source, uv_interp + pixel_size * vec2(1.0, -1.0), 0.0);
+	vec4 D = textureLod(source, uv_interp + pixel_size * vec2(-0.5, -0.5), 0.0);
+	vec4 E = textureLod(source, uv_interp + pixel_size * vec2(0.5, -0.5), 0.0);
+	vec4 F = textureLod(source, uv_interp + pixel_size * vec2(-1.0, 0.0), 0.0);
+	vec4 G = textureLod(source, uv_interp, 0.0);
+	vec4 H = textureLod(source, uv_interp + pixel_size * vec2(1.0, 0.0), 0.0);
+	vec4 I = textureLod(source, uv_interp + pixel_size * vec2(-0.5, 0.5), 0.0);
+	vec4 J = textureLod(source, uv_interp + pixel_size * vec2(0.5, 0.5), 0.0);
+	vec4 K = textureLod(source, uv_interp + pixel_size * vec2(-1.0, 1.0), 0.0);
+	vec4 L = textureLod(source, uv_interp + pixel_size * vec2(0.0, 1.0), 0.0);
+	vec4 M = textureLod(source, uv_interp + pixel_size * vec2(1.0, 1.0), 0.0);
+
+	float weight = 0.5 / 4.0;
+	float lesser_weight = 0.125 / 4.0;
+
+	frag_color = (D + E + I + J) * weight;
+	frag_color += (A + B + G + F) * lesser_weight;
+	frag_color += (B + C + H + G) * lesser_weight;
+	frag_color += (F + G + L + K) * lesser_weight;
+	frag_color += (G + H + M + L) * lesser_weight;
+#endif
 }

drivers/gles3/storage/texture_storage.cpp

@@ -1837,10 +1837,10 @@ void TextureStorage::_create_render_target_backbuffer(RenderTarget *rt) {
 		}
 		GLES3::Utilities::get_singleton()->texture_allocated_data(rt->backbuffer, texture_size_bytes, "Render target backbuffer color texture");
 
-		// Initialize all levels to opaque Magenta.
+		// Initialize all levels to clear black.
 		for (int j = 0; j < count; j++) {
 			glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, rt->backbuffer, j);
-			glClearColor(1.0, 0.0, 1.0, 1.0);
+			glClearColor(0.0, 0.0, 0.0, 0.0);
 			glClear(GL_COLOR_BUFFER_BIT);
 		}
 
@@ -2568,18 +2568,18 @@ void TextureStorage::render_target_copy_to_back_buffer(RID p_render_target, cons
 	}
 
 	glDisable(GL_BLEND);
-	//single texture copy for backbuffer
+	// Single texture copy for backbuffer.
 	glBindFramebuffer(GL_FRAMEBUFFER, rt->backbuffer_fbo);
 	glActiveTexture(GL_TEXTURE0);
 	glBindTexture(GL_TEXTURE_2D, rt->color);
 	GLES3::CopyEffects::get_singleton()->copy_screen();
 
 	if (p_gen_mipmaps) {
-		GLES3::CopyEffects::get_singleton()->bilinear_blur(rt->backbuffer, rt->mipmap_count, region);
+		GLES3::CopyEffects::get_singleton()->gaussian_blur(rt->backbuffer, rt->mipmap_count, region, rt->size);
 		glBindFramebuffer(GL_FRAMEBUFFER, rt->backbuffer_fbo);
 	}
 
-	glEnable(GL_BLEND); // 2D almost always uses blend.
+	glEnable(GL_BLEND); // 2D starts with blend enabled.
 }
 
 void TextureStorage::render_target_clear_back_buffer(RID p_render_target, const Rect2i &p_region, const Color &p_color) {
@@ -2624,8 +2624,10 @@ void TextureStorage::render_target_gen_back_buffer_mipmaps(RID p_render_target,
 			return; //nothing to do
 		}
 	}
+	glDisable(GL_BLEND);
+	GLES3::CopyEffects::get_singleton()->gaussian_blur(rt->backbuffer, rt->mipmap_count, region, rt->size);
+	glEnable(GL_BLEND); // 2D starts with blend enabled.
 
-	GLES3::CopyEffects::get_singleton()->bilinear_blur(rt->backbuffer, rt->mipmap_count, region);
 	glBindFramebuffer(GL_FRAMEBUFFER, rt->backbuffer_fbo);
 }